Optimization of the parameters of a method for computer-aided detection of perfusion deficiencies in brain images

Abstract

Objective: Simulated data from the recent Institute of Physics and Engineering in Medicine audit of quantitative cerebral perfusion were used to optimize the parameters of eigenimage analysis, a method for computer-aided detection.

Methods: Twenty normal images provided by the audit were registered to the International Consortium for Brain Mapping 452 template using HERMES multimodality software and normalized to total counts. Six normal atlases were formed using the mean image and from zero to five eigenimages. Eight patient images, with computer-simulated lesions at known positions, were similarly registered and normalized. For each atlas, z-score images were formed for each patient. Thresholds of z0 = 2-5 in intervals of 0.5 were applied to the z-score images to form binary images of normal and abnormal voxels. A lesion was defined as a connected group of abnormal voxels with a minimum size of 1 ml. Lesions were assigned to one of 12 regions defined by the template. For eight patients, this gave 96 regions, 19 of which were known to contain an abnormality. Receiver-operating characteristic analysis was performed for the regions using z0 as a variable threshold.

Results: For the receiver-operating characteristic analysis, an optimal area under the curve of approximately 0.90 was found using either one or three eigenimages, whereas good results (sensitivity = 0.75%; specificity = 90%) were obtained for a threshold of z0 approximately equal to 3. When the number of images in the normal dataset was considered, a meta-analysis showed consistency with other studies.

Conclusion: Eigenimage analysis was shown to give good diagnostic accuracy for cerebral perfusion images based on objective evaluation using simulated images.